A Robust Structured Light Pattern Decoding Method for Single-Shot 3D Reconstruction

Traditional pattern element identification methods in binary shape-coded structured light are usually lack of robustness to the surface colors and textures. This paper introduces a novel pattern decoding method for a binary structured light pattern, which is composed of eight geometrical elements. The pattern elements are designed as grid shape and the intersection of grid lines is defined as the feature point. By extracting the grid-points firstly, a topological network is constructed to separate each pattern element from the image. Then, pattern element identification is modeled as a supervised classification problem. The convolutional neural network (CNN) is applied to classify the pattern elements. The network is trained with a mass of pattern element samples with various blur and distortion. The experimental results show that the proposed pattern element identification method has strong robustness to surface color, texture, distortion and image noise.